Abstract
Purpose :
Peripapillary retinal blood vessels (RBVs) are currently segmented within the retinal nerve fiber layer (RNFL) in commercially available OCT segmentation software, contributing to automated calculation of RNFL thickness (std-RNFLT). In this cross-sectional clinical study we developed a fully automated segmentation algorithm capable of subtracting RBVs from RNFLT using OCT Angiography (OCT-A) scans of the optic nerve head (ONH), and we compared vessel-free RNFLT (vf-RNFLT) to std-RNFLT in order to evaluate the impact of RBVs on RNFLT.
Methods :
ONH scans of 22 healthy eyes and 4 explicative glaucomatous eyes were analysed. After convolutional neural network-based RNFL and disc margin segmentation, an en face OCT-A image of the RNFL plexus was generated, enhanced and thresholded. Assuming vessels’ tubularity, the RBV thickness in the axial direction was extrapolated from vessel pixels distance to the nearest non-vessel pixel. The contribution of RBV thickness was then subtracted from the std-RNFLT map to obtain vf-RNFLT. Std- and vf-RNFLT of healthy eyes were compared with t-test. The influence of covariates was analysed by a linear mixed-effects model (GLMM).
Results :
In healthy subjects, average (Avg), superior (S), nasal (N), inferior (I), temporal (T) std-RNFLT and vf-RNFLT (all values in microns, mean±SD) were 111.6±25.1 vs. 104.1±21.7 (p = .035), 134.7±14.2 vs. 124.2±13.6 (p = .016), 82.9±11.3 vs. 78.2±10.5 (p = .16), 130±14.1 vs. 117±12.3 (p = .002), 98.7±11.7 vs. 96.9±12 (p = .633). The clock hour thickness analysis showed that clock hours 5, 7, 12 were significantly different (all p < .05). Peripapillary RBVs accounted for 6.3%, 7.8%, 5.6%, 10%, and 1.8%, of the Avg, S, N, I, T std-RNFLT in healthy eyes, and 9,8%, 15%, 9.4%, 12.7%, and 2.1% in glaucomatous eyes. GLMM showed no association of age, axial length and keratometry values with the difference in clock hour std- and vf-RNFLT.
Conclusions :
Peripapillary RBVs account for a significant percentage of the RNFLT, if measured by commercially available OCT segmentation software. Our fully automated deep learning-based segmentation software excluding the RBVs, which are spared by glaucomatous damage, provides more accurate estimates of RNFLT, possibly improving OCT ability to detect early neural damage due to glaucoma.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.